A smart distribution room operation and maintenance method and system based on multi-source data fusion

Abstract

The present invention relates to a smart distribution room operation and maintenance method and system based on multi-source data fusion. The method comprises: collecting distribution room environment and equipment data through multi-source sensing equipment to obtain original multi-source data; standardizing the original multi-source data to obtain standardized heterogeneous data; modeling the spatiotemporal correlation between equipment vibration characteristics and equipment power parameters through a spatiotemporal graph convolutional network, and fusing cross-modal deep features extracted from standardized heterogeneous data to obtain panoramic perception features; based on the panoramic perception features, performing anomaly detection through an isolation forest-long short-term memory hybrid model, and deploying a causal reasoning engine to analyze the causal relationship between multiple variables to obtain a distribution room health status assessment result; making operation and maintenance decisions and performing operation and maintenance on the distribution room based on the distribution room health status assessment result. The present invention realizes deep fusion of multi-source data, can accurately detect equipment anomalies and analyze the root causes, and improve the efficiency and reliability of distribution room operation and maintenance.

Description

Technical Field

[0001] The present invention relates to the technical field of power grid management, and in particular to a smart distribution room operation and maintenance method and system based on multi-source data fusion. Background Art

[0002] As power systems become increasingly intelligent, the safe and stable operation of distribution rooms, key nodes in power transmission and distribution, is crucial to the entire system. Traditional distribution room operations and maintenance rely primarily on regular manual inspections and single-sensor monitoring. These issues include limited data collection, inadequate data utilization, a lack of device correlation analysis, limited anomaly detection capabilities, and a lack of root-cause analysis capabilities. While intelligent sensing technology and data analysis methods have made significant progress in recent years, distribution room operations and maintenance still face numerous technical challenges.

[0003] Existing data fusion technologies for power distribution rooms mainly use simple splicing or weighted averaging to process multi-source heterogeneous data, which makes it difficult to accurately reflect the actual operating status of the equipment. In addition, existing technologies are mostly based on correlation analysis rather than causal reasoning, lacking automated root cause analysis capabilities and unable to determine the source and propagation path of anomalies, increasing the difficulty and time cost of troubleshooting. Summary of the Invention

[0004] The present invention provides a smart distribution room operation and maintenance method and system based on multi-source data fusion to address the shortcomings of the existing technology.

[0005] The present invention provides a smart distribution room operation and maintenance method based on multi-source data fusion, comprising:

[0006] S1: Collect distribution room environment and equipment data through multi-source sensing equipment to obtain original multi-source data;

[0007] S2: performing standardization processing on the original multi-source data to obtain standardized heterogeneous data;

[0008] S3: The spatiotemporal correlation between the equipment vibration characteristics and the equipment power parameters is modeled through a spatiotemporal graph convolutional network, and the cross-modal deep features extracted from the standardized heterogeneous data are integrated to obtain panoramic perception features;

[0009] S4: Based on the panoramic perception features, anomaly detection is performed using an isolation forest-long short-term memory hybrid model. A causal inference engine is deployed to analyze the causal relationship between multiple variables to obtain a health status assessment result for the distribution room.

[0010] S5: Making an operation and maintenance decision based on the health status assessment result of the power distribution room, and performing operation and maintenance on the power distribution room according to the operation and maintenance decision.

[0011] According to a smart distribution room operation and maintenance method based on multi-source data fusion provided by the present invention, the original multi-source data in step S1 includes:

[0012] Ambient temperature and humidity data, which is collected by a temperature and humidity sensor;

[0013] Gas concentration data, the gas concentration data is collected by a gas concentration sensor;

[0014] Power operation parameter data, the power operation parameter data is collected by a power parameter collector;

[0015] Equipment vibration characteristic data, the equipment vibration characteristic data is collected by a vibration sensor;

[0016] Thermal imaging data, which is collected by infrared thermal imaging equipment.

[0017] According to a smart distribution room operation and maintenance method based on multi-source data fusion provided by the present invention, step S2 specifically includes:

[0018] Performing noise removal and outlier detection on the original multi-source data using a data cleaning algorithm to obtain cleaned data;

[0019] Normalizing the numerical data in the cleaned data to the maximum and minimum values through normalization processing to obtain normalized numerical data;

[0020] Performing a logarithmic transformation on the skewed distribution data in the cleaned data to obtain balanced distribution data;

[0021] The time synchronization mechanism is used to align the timestamps of data with different sampling frequencies to obtain time-series consistent data;

[0022] Completing the missing parts in the time series consistent data by using a missing value filling algorithm to obtain integrity-enhanced data;

[0023] The integrity enhanced data is converted into a unified format through a data format converter to obtain the standardized heterogeneous data.

[0024] According to a smart distribution room operation and maintenance method based on multi-source data fusion provided by the present invention, step S3 further includes:

[0025] S31: Establishing a device topology map based on the device connection relationship in the power distribution room, and adjusting the edge weights of the device topology map according to the device operation status data to obtain a dynamic adjacency matrix;

[0026] S32: Inputting the dynamic adjacency matrix into the spatiotemporal graph convolutional network to obtain a spatiotemporal joint feature vector;

[0027] S33: performing attention interaction calculation on the device power parameter and the spatiotemporal joint feature vector to obtain a coupling coefficient matrix, and performing abnormal region detection based on a threshold segmentation method on the coupling coefficient matrix to obtain coupling features;

[0028] S34: performing feature extraction on the standardized heterogeneous data to obtain cross-modal deep features;

[0029] S35: Fusing the coupling feature and the cross-modal depth feature to obtain a panoramic perception feature.

[0030] According to a smart distribution room operation and maintenance method based on multi-source data fusion provided by the present invention, step S32 further includes:

[0031] S321: performing spectral graph convolution on the dynamic adjacency matrix through Chebyshev polynomial approximation to extract spatial correlation features of the equipment vibration spectrum;

[0032] S322: Inputting the spatial correlation features into a gated temporal convolutional network to extract temporal correlation features of the vibration signal time series;

[0033] S323: combining the spatial correlation features and the temporal correlation features to obtain fusion features;

[0034] S324: Add the fused features to the original vibration signal in the standardized heterogeneous data through residual connection to obtain a spatiotemporal joint feature vector.

[0035] According to a smart distribution room operation and maintenance method based on multi-source data fusion provided by the present invention, step S4 further includes:

[0036] S41: Inputting the panoramic perception features into the isolation forest-long short-term memory hybrid model to obtain device anomaly determination results including time series anomalies and sudden anomalies;

[0037] S42: locating the root cause of the abnormal device in the device abnormality determination result by using a structural equation model, and obtaining a distribution room health status assessment result including the device abnormality determination result and the root cause analysis result.

[0038] According to a smart distribution room operation and maintenance method based on multi-source data fusion provided by the present invention, step S41 specifically includes:

[0039] S411: Calculating prediction residuals using the LSTM branch of the isolation forest-long short-term memory hybrid model, marking devices whose prediction residuals are greater than a first preset threshold as abnormal, and obtaining a first marking result;

[0040] The expression of the first preset threshold is: ,in is the prediction residual, is the standard deviation of the residual series;

[0041] S412: Calculating a path length score using the isolation forest branch of the isolation forest-long short-term memory hybrid model, and marking devices with a path length score greater than a second preset threshold as abnormal to obtain a second marking result;

[0042] The expression of the second threshold is: ,in Score path length;

[0043] S413: When either the first marking result or the second marking result is marked as abnormal, the device abnormality determination result is output as abnormal; when either the first marking result or the second marking result is marked as no abnormality, the device abnormality determination result is output as normal.

[0044] According to a smart distribution room operation and maintenance method based on multi-source data fusion provided by the present invention, step S42 further includes:

[0045] S421: Define latent variables and observed variables, describe the variable relationship between latent variables and observed variables, and construct a structural equation model;

[0046] S422: Calculating the path coefficients between variables of the structural equation model by partial least squares path analysis;

[0047] S423: performing root cause verification using a counterfactual intervention method based on the structural equation model and the path coefficients between the variables to obtain root cause analysis results;

[0048] S424: Output the equipment abnormality determination result and the root cause analysis result to obtain a health status assessment result of the distribution room.

[0049] According to a smart distribution room operation and maintenance method based on multi-source data fusion provided by the present invention, the expression of the variable relationship in step S421 is:

[0050]

[0051] in, Indicates the RMS value of vibration, Indicates mechanical wear, Indicates the SF6 gas concentration, Indicates insulation degradation, Indicates the amount of partial discharge, Indicates environmental corrosion, represents the rate of temperature rise, Indicates the Factor loading coefficients, Indicates the measurement error terms;

[0052] The expression of the structural equation model in step S421 is:

[0053] in, Indicates the path coefficients, Indicates the structural residual terms;

[0054] The expression of the path coefficient between the variables in step S422 is: in, represents the estimated path coefficient between variables, Indicates the total number of paths, Indicates the Insulation degradation caused by the Indicates the Mechanical wear caused by the path, Indicates the Environmental corrosion caused by a path.

[0055] The present invention also provides a smart distribution room operation and maintenance system based on multi-source data fusion, comprising:

[0056] Acquisition module: used to collect distribution room environment and equipment data based on multi-source sensing equipment to obtain original multi-source data;

[0057] Standardization module: used for standardizing the original multi-source data to obtain standardized heterogeneous data;

[0058] Feature extraction module: used to perform spatiotemporal correlation modeling of equipment vibration characteristics and equipment power parameters through a spatiotemporal graph convolutional network, and to fuse the cross-modal deep features extracted from the standardized heterogeneous data to obtain panoramic perception features;

[0059] Detection module: used to detect anomalies using an isolation forest-long short-term memory hybrid model based on the panoramic perception features, and deploy a causal inference engine to analyze the causal relationship between multiple variables to obtain a health status assessment result of the distribution room;

[0060] Decision output module: used to make operation and maintenance decisions based on the health status assessment results of the distribution room, and operate and maintain the distribution room based on the operation and maintenance decisions.

[0061] The present invention provides an intelligent distribution room operation and maintenance method and system based on multi-source data fusion. By combining a spatiotemporal graph convolutional network with a multimodal fusion algorithm, the system achieves deep fusion of multi-source heterogeneous data in the distribution room. It can not only process different types of data such as temperature and humidity, gas concentration, power parameters, equipment vibration and thermal imaging, but also simultaneously capture the spatial topological relationship between devices and the time evolution characteristics of parameters. In particular, the present invention uses Chebyshev polynomials for spectral graph convolution, combined with a gated time convolutional network and residual connection, to effectively extract the complex spatiotemporal coupling relationship between equipment vibration and power parameters, greatly improving the comprehensive perception capability of the operating status of the distribution room, avoiding the data island problem in traditional methods, and providing a more comprehensive and accurate data basis for anomaly detection and root cause analysis. The present invention also innovatively proposes an isolation forest-long short-term memory hybrid model, which organically combines the distance-based isolation forest algorithm with the prediction-based LSTM model, and constructs a dual detection mechanism that can simultaneously detect sudden anomalies and timing anomalies. The LSTM branch calculates the prediction residual to identify timing anomalies, and the isolation forest branch calculates the path length score to identify sudden anomalies. The two complement each other's advantages, greatly improving the accuracy and coverage of anomaly detection. In particular, for the two typical anomalies common in power equipment - timing anomalies caused by slow degradation and sudden anomalies caused by sudden failures, the hybrid model of the present invention shows excellent detection performance, greatly reducing the false alarm rate and missed alarm rate, and providing a reliable basis for operation and maintenance decisions. In addition, the present invention uses structural equation models and counterfactual intervention methods for root cause analysis, breaking through the limitations of traditional correlation analysis, and defining latent variables. and observed variables and describe the relationship between them, constructing a causal network of the equipment status in the distribution room, which can accurately identify the root cause and propagation path of the anomaly. In practical applications, this root cause analysis method based on causal reasoning can quickly locate the source of the fault, greatly improve maintenance efficiency, reduce downtime, and reduce maintenance costs. The method of the present invention can not only effectively process static data, but also establish a dynamic adjacency matrix update mechanism based on the operating status of the equipment, which can adjust the edge weights of the equipment topology graph in real time, so that the model can adaptively reflect changes in the equipment status. The dynamic adaptive mechanism makes the system have good adaptability to long-term evolution factors such as equipment aging and environmental changes, and maintains long-term stable performance. At the same time, the coupling coefficient matrix obtained by attention interaction calculation provides accurate positioning of the abnormal area, further improving the adaptability and accuracy of the system.

[0062] The present invention realizes intelligent operation and maintenance decision-making based on the health status assessment results. By combining the equipment abnormality judgment results with the root cause analysis results, a complete information chain is provided for operation and maintenance personnel, which greatly reduces the cognitive burden of operation and maintenance personnel, improves decision-making efficiency and accuracy, realizes proactive prevention work in distribution room operation and maintenance, effectively extends the service life of equipment, and improves the overall reliability of the distribution system. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to more clearly illustrate the technical solutions in the present invention or the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0064] Figure 1 A flowchart of a smart distribution room operation and maintenance method based on multi-source data fusion provided by an embodiment of the present invention;

[0065] Figure 2 A schematic diagram of the structure of an intelligent distribution room operation and maintenance system based on multi-source data fusion provided in an embodiment of the present invention.

[0066] Figure numerals: 100, acquisition module; 200, standardization module; 300, feature extraction module; 400, detection module; 500, decision output module. DETAILED DESCRIPTION

[0067] In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the drawings in the present invention. Obviously, the embodiments described are part of the embodiments of the present invention, not all of the embodiments, and they should not be understood as limitations on the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. In the description of the present invention, it should be understood that the terms used are only for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0068] The following describes embodiments of the present invention with reference to the accompanying drawings.

[0069] Such as Figure 1 As shown, the present invention provides a smart distribution room operation and maintenance method based on multi-source data fusion, including:

[0070] S1: Collect distribution room environment and equipment data through multi-source sensing equipment to obtain original multi-source data.

[0071] Among them, the original multi-source data in step S1 includes: ambient temperature and humidity data, which are collected by temperature and humidity sensors; gas concentration data, which are collected by gas concentration sensors; power operation parameter data, which are collected by power parameter collectors; equipment vibration characteristic data, which are collected by vibration sensors; thermal imaging data, which are collected by infrared thermal imaging equipment.

[0072] Furthermore, ambient temperature and humidity data is collected using temperature and humidity sensors placed in key areas of the distribution room, such as around transformers, inside switchgear, and near ventilation openings, to monitor temperature and humidity changes in different areas in real time. Gas concentration data is collected using gas concentration sensors, which primarily monitor SF6, ozone, carbon monoxide, and other gas components within the distribution room. These sensors use electrochemical sensing technology or infrared absorption spectroscopy to selectively detect multiple gases, with sensitivity reaching the ppm level. Power operation parameter data is collected using power parameter collectors, which are directly connected to key nodes in the distribution system and monitor core electrical parameters such as voltage, current, power factor, and harmonic content. Equipment vibration characteristic data is collected using vibration sensors installed on key equipment such as transformers, circuit breakers, and switchgear, monitoring parameters such as vibration amplitude, frequency, and acceleration. Thermal imaging data is collected using infrared thermal imaging equipment, which performs contactless scanning of the surface temperature distribution of equipment within the distribution room, generating thermal images.

[0073] S2: performing standardization processing on the original multi-source data to obtain standardized heterogeneous data.

[0074] Among them, step S2 specifically includes: removing noise and detecting outliers on the original multi-source data through a data cleaning algorithm to obtain cleaned data; normalizing the maximum and minimum values of the numerical data in the cleaned data through normalization processing to obtain normalized numerical data; transforming the skewed distribution data in the cleaned data through logarithmic transformation to obtain balanced distribution data; aligning the timestamps of data with different sampling frequencies through a time synchronization mechanism to obtain time-series consistent data; completing the missing parts in the time-series consistent data through a missing value filling algorithm to obtain integrity-enhanced data; converting the integrity-enhanced data into a unified format through a data format converter to obtain the standardized heterogeneous data.

[0075] Furthermore, the standardized heterogeneous data obtained through preprocessing significantly improved data quality and availability. First, median filtering and Z-score methods were used to remove noise and detect outliers from the original data, effectively eliminating data contamination caused by factors such as electromagnetic interference and sensor failure. Second, different types of data were standardized through maximum and minimum value normalization and logarithmic transformation, solving the dimensional inconsistency and uneven distribution problems of multi-source heterogeneous data. Third, a time synchronization mechanism based on UTC timestamps and a multi-strategy missing value filling algorithm were used to overcome technical difficulties such as large differences in sampling frequencies of different sensors and missing data. Finally, a unified data format and descriptor system was constructed, laying a solid foundation for subsequent spatiotemporal correlation modeling and multimodal fusion. Compared with traditional simple data cleaning methods, this preprocessing method not only retains the key features of the original data, but also enhances the comparability and correlation between data through standardized transformation, thereby providing high-quality data support for anomaly detection and fault diagnosis in smart distribution rooms.

[0076] S3: The spatiotemporal correlation between the equipment vibration characteristics and the equipment power parameters is modeled through a spatiotemporal graph convolutional network, and the cross-modal deep features extracted from the standardized heterogeneous data are integrated to obtain panoramic perception features.

[0077] Wherein, step S3 further includes:

[0078] S31: Based on the connection relationship of the devices in the power distribution room, a device topology map is established, and the edge weights of the device topology map are adjusted according to the operating status data of the devices to obtain a dynamic adjacency matrix.

[0079] In step S31, a device topology is first established based on the device connectivity within the distribution room. The edge weights of the device topology are adjusted based on the device's operating status data to obtain a dynamic adjacency matrix. The device topology is a graph-theoretic representation of the physical connections and electrical relationships between the devices within the distribution room. Each node represents a device (such as a transformer, switchgear, or circuit breaker), and the edges between nodes represent the physical or electrical connections between the devices.

[0080] The dynamic adjacency matrix is a mathematical representation of the device topology. Matrix elements represent the strength of the connection between two devices. The initial adjacency matrix is determined based on the physical connections between the devices. Edge weights are then dynamically adjusted based on device operating status data. Weight adjustment utilizes a state vector composed of device vibration characteristics and power parameters. When two devices have similar states, the edge weight between them increases; otherwise, it decreases. This dynamically adjusted adjacency matrix reflects the real-time strength of connections between devices, based not only on physical connections but also on their operating states.

[0081] S32: Input the dynamic adjacency matrix into the spatiotemporal graph convolutional network to obtain a spatiotemporal joint feature vector.

[0082] Furthermore, in step S32, the dynamic adjacency matrix is input into a spatiotemporal graph convolutional network to obtain a spatiotemporal joint feature vector. The spatiotemporal graph convolutional network is a deep learning architecture that combines a graph convolutional network (GCN) and a temporal convolutional network (TCN), which can simultaneously capture the spatial dependency and temporal dynamics of data.

[0083] Wherein, step S32 further includes:

[0084] S321: Performing spectral graph convolution on the dynamic adjacency matrix through Chebyshev polynomial approximation to extract spatial correlation features of the equipment vibration spectrum.

[0085] First, in step S321, the dynamic adjacency matrix is convolved with the Chebyshev polynomial approximation to extract the spatial correlation characteristics of the equipment vibration spectrum. The Chebyshev polynomial approximation is a method for quickly calculating the spectral convolution, avoiding the complex operation of directly calculating the eigendecomposition of the graph Laplace matrix. Specifically, the K-order Chebyshev polynomial approximation formula is: ,in, is the extracted spatial correlation feature, is the polynomial order, is the learnable coefficient, is the Chebyshev polynomial, is the vibration signal time series. In practical applications, A good approximation effect can be obtained. Through the approximation operation, the vibration spectrum characteristics of each device node can be weightedly combined with the characteristics of its adjacent devices to extract spatial correlation features, reflecting the mutual influence and correlation pattern between devices.

[0086] S322: Input the spatial correlation features into a gated temporal convolutional network to extract the temporal correlation features of the vibration signal time series.

[0087] In step S33, the spatial correlation features in step S321 are input into the gated time convolutional network to extract the time correlation features of the vibration signal time series. The gated time convolutional network is an improved one-dimensional convolutional neural network that introduces a gating mechanism to control the flow of information. The gating mechanism is similar to the gating unit in the long short-term memory network (LSTM), which can selectively retain or ignore time series information and effectively process long-term dependencies in long sequence data. In the present invention, the gated time convolutional network processes the time series vibration data of each device node, and applies a multi-layer convolution structure, each layer containing convolution kernels of different sizes to capture pattern changes at different time scales. Residual connections and batch normalization layers are also set between the convolution layers to accelerate the training process and avoid the gradient vanishing problem. Through the gated time convolutional network, the dynamic characteristics of the device vibration signal that change over time are extracted, such as the periodicity, trend and mutation characteristics of the vibration pattern.

[0088] S323: Concatenate the spatial correlation features and the temporal correlation features to obtain fusion features.

[0089] Next, the spatially correlated features and temporally correlated features are concatenated to obtain fused features. The feature concatenation operation connects the spatially correlated features and temporally correlated features in the feature dimension, producing a higher-dimensional feature vector. This concatenation operation preserves the complete information of both spatial and temporal features. Specifically, during concatenation, the output features of the spatial graph convolution and the output features of the temporal convolution are concatenated along the channel dimension and fed into the spatiotemporal feature fusion layer. Cross-modal feature fusion is achieved through 1×1 convolution. Through this transformation, the network can learn the complex relationship between spatial and temporal features, generating more expressive fused features.

[0090] S324: Add the fused features to the original vibration signal in the standardized heterogeneous data through residual connection to obtain a spatiotemporal joint feature vector.

[0091] After obtaining the fused features, a residual connection mechanism is introduced in step S324 to add the original vibration signal to the fused features, alleviate the gradient vanishing problem, and generate a spatiotemporal joint feature vector. Residual connection is a commonly used technique in deep neural networks. By adding cross-layer connections in the forward propagation, it effectively solves the gradient vanishing problem of deep networks. Residual connection transmits the original signal information directly to the deep network, ensuring that important basic features are not lost in multi-layer processing. At the same time, it allows the network to learn the residual mapping between the original signal and the processed features, making it easier to optimize and train. Through residual connection, the spatiotemporal joint feature vector not only contains the basic features of the original vibration signal, but also integrates the high-order features extracted by the deep network, forming a more comprehensive and robust feature representation.

[0092] S33: Perform attention interaction calculation on the device power parameter and the spatiotemporal joint feature vector to obtain a coupling coefficient matrix, and perform abnormal area detection based on a threshold segmentation method on the coupling coefficient matrix to obtain coupling features.

[0093] In step S33, a cross-modal cross-attention calculation is performed on the spatiotemporal joint feature vector and the power parameters (current and voltage harmonics) to generate a coupling coefficient matrix between the mechanical vibration of the equipment and the electrical parameters; subsequently, an adaptive threshold segmentation algorithm is used to detect abnormal areas on the coupling coefficient matrix to obtain a mechanical-electrical coupling abnormal pattern identifier.

[0094] Specifically, first, attention interaction calculation is performed on the device power parameters and the spatiotemporal joint feature vector. Attention interaction calculation is a feature interaction method based on the attention mechanism, which is used to capture the relationship between different feature domains. In the present invention, attention interaction calculation is used to model the relationship between the device power parameters and vibration characteristics. The specific calculation process is as follows: First, the power parameter vector and the spatiotemporal joint feature vector are mapped to the query space, key space and value space respectively through linear transformation; then, the attention weight is calculated, and the attention weight represents the attention degree of each power parameter to each vibration feature; finally, the value vector is weighted and summed by the attention weight to obtain the attention output of the power parameter to the vibration feature, and the coupling coefficient matrix is obtained by recalculating the correlation between the power parameter and the attention output, which characterizes the coupling strength between the i-th power parameter and the j-th vibration feature.

[0095] The subsequent abnormal region detection based on the threshold segmentation method of the coupling coefficient matrix is to find the area where the coupling coefficient significantly deviates from the normal mode. The threshold segmentation method uses an adaptive threshold: ,in is the threshold value obtained, is the mean of the coupling coefficient matrix, is the standard deviation of the coupling coefficient matrix, It is an adjustable parameter, usually set to 2. When the coupling coefficient is greater than When , it is determined to be an abnormal area, indicating that there is an abnormally strong coupling relationship between the corresponding power parameters and vibration characteristics, which may indicate equipment failure. After threshold segmentation, the coupling coefficient of the abnormal area is retained and the other areas are set to zero, forming a sparse coupling feature matrix, which highlights the potential abnormal pattern.

[0096] S34: Perform feature extraction on the standardized heterogeneous data to obtain cross-modal deep features.

[0097] Step S34 performs feature extraction on the standardized heterogeneous data to obtain cross-modal deep features. Cross-modal deep feature extraction is to extract semantically consistent and complementary feature representations from multiple different types of data. In the distribution room operation and maintenance scenario, in addition to equipment vibration and power parameters, it also includes multimodal information such as temperature and humidity data, gas concentration data, and thermal imaging data.

[0098] This invention uses a combination of modality-specific encoders and shared encoders to extract features from these heterogeneous data. The modality-specific encoders design different feature extraction networks for each data type: a multi-layer perceptron (MLP) is used to extract environmental features for temperature and humidity data; a one-dimensional convolutional neural network (1D CNN) is used to extract gas features for gas concentration data; and a two-dimensional convolutional neural network (2D CNN) is used to extract thermal distribution features for thermal imaging data. Each dedicated encoder maps the raw data into a latent space, while the shared encoder uses techniques such as adversarial training and mutual information maximization to map features from different modalities into a common semantic space. Subsequent adversarial training is then used to map features from different modalities back into the same distribution space. By combining these modality-specific and shared encoders, this invention extracts semantically consistent cross-modal deep features from multimodal data. These features describe the state of the distribution room from different perspectives.

[0099] S35: Fusing the coupling feature and the cross-modal depth feature to obtain a panoramic perception feature.

[0100] In step S35, the present invention fuses the coupled features and cross-modal deep features, with the aim of integrating features from different sources and in different representations into a unified, more expressive feature representation. The feature fusion in the present invention adopts a multi-level attention fusion mechanism, including feature-level attention and decision-level attention. Feature-level attention calculates the importance weight of each feature, and through feature-level attention, assigns different weights to each element in the coupled features and cross-modal deep features to highlight important information; decision-level attention calculates the overall weight of the coupled features and cross-modal deep features, and the final panoramic perception feature is obtained through weighted fusion.

[0101] S4: Based on the panoramic perception features, anomaly detection is performed through the isolation forest-long short-term memory hybrid model, and a causal reasoning engine is deployed to analyze the causal relationship between multiple variables to obtain the health status assessment results of the distribution room.

[0102] Wherein, step S4 further includes:

[0103] S41: Inputting the panoramic perception features into an isolation forest-long short-term memory hybrid model to obtain device anomaly determination results including time series anomalies and sudden anomalies.

[0104] In step S41, the panoramic perception features are first input into the Isolation Forest-Long Short-Term Memory hybrid model to obtain device anomaly determination results, including both sequential and sudden anomalies. The Isolation Forest-Long Short-Term Memory hybrid model provided by this invention is an innovative anomaly detection architecture that combines the isolation-based Isolation Forest algorithm with the prediction-based Long Short-Term Memory (LSTM) network. It can simultaneously identify sudden and sequential anomalies in data. The hybrid model contains two parallel branches: the LSTM branch is responsible for detecting sequential anomalies, and the Isolation Forest branch is responsible for detecting sudden anomalies.

[0105] Wherein, step S41 specifically includes:

[0106] S411: Calculate the prediction residual using the LSTM branch of the Isolation Forest-Long Short-Term Memory hybrid model, and mark devices whose prediction residual is greater than a first preset threshold as abnormal to obtain a first marking result; the expression of the first preset threshold is: ,in is the prediction residual, is the standard deviation of the residual series.

[0107] In the LSTM branch, the present invention takes the panoramic perception features of the past N time steps as input, predicts the feature value of the next time step, and presets For the LSTM network's prediction objective, the network is trained by minimizing the mean squared error between the predicted and actual values. After training, the present invention predicts new data and calculates the prediction residual, or prediction error. This is the Euclidean distance between the actual observed value and the predicted value. The prediction residual reflects the degree to which the current observation deviates from the model's prediction. Large residual values generally indicate that the data contain unusual patterns.

[0108] Then, the devices whose prediction residuals are greater than the first preset threshold are marked as abnormal to obtain the first marking result. The threshold setting method is the 3σ principle. Under the normal distribution assumption, the probability of data falling outside the range of μ±3σ is only 0.3%. Therefore, data points outside this range are regarded as abnormal. Therefore, the present invention compares the prediction residuals of each time point with the threshold size, marks the abnormal time points as abnormal, and forms a preliminary time series anomaly detection result, namely the first marking result.

[0109] S412: Calculate the path length score using the isolation forest branch of the isolation forest-long short-term memory hybrid model, and mark devices with the path length score greater than a second preset threshold as abnormal to obtain a second marking result; the expression of the second threshold is: in Score the path length.

[0110] Furthermore, Isolation Forest is a tree-based anomaly detection algorithm. It constructs a forest of multiple isolation trees (iTrees) to calculate the difficulty of isolating a data point. The isolation tree construction process randomly selects a dimension in the feature space, randomly chooses a split point on that dimension, and then divides the data into two parts. This process is recursively repeated for each split until the tree height limit is reached or the node contains only one data point. For a given data point, the path length from the root node to the leaf node in each isolation tree is calculated, and the average of all trees is taken as the anomaly score for that point.

[0111] In the Isolation Forest Algorithm, path length score The value range of is (0,1). The closer to 1, the more abnormal the data point is. Usually, in the isolation forest algorithm, data points with a path length score greater than 0.8 are considered abnormal. By comparing the path length score of each data point with the threshold, the abnormal data points are marked as abnormal, forming a sudden anomaly detection result, that is, the second marking result.

[0112] S413: When either the first marking result or the second marking result is marked as abnormal, the device abnormality determination result is output as abnormal; when either the first marking result or the second marking result is marked as no abnormality, the device abnormality determination result is output as normal.

[0113] In step S413, as long as the data point is identified as abnormal by any branch, the final judgment result is abnormal. The abnormal reporting strategy of the present invention is suitable for high requirements on the sensitivity of abnormal detection, such as early fault warning of distribution equipment, and it is better to have a false alarm than to miss potential risks.

[0114] S42: locating the root cause of the abnormal device in the device abnormality determination result by using a structural equation model, and obtaining a distribution room health status assessment result including the device abnormality determination result and the root cause analysis result.

[0115] In step S42, the present invention uses a structural equation model to locate the root cause of abnormal devices in the device abnormality determination results, obtaining a distribution room health status assessment result that includes the device abnormality determination results and the root cause analysis results. The Structural Equation Model (SEM) is a statistical method that can analyze causal relationships between multiple variables. It combines factor analysis and path analysis and is suitable for modeling complex relationships between latent variables (variables that cannot be directly observed) and observed variables.

[0116] Wherein, step S42 further includes:

[0117] S421: Define latent variables and observed variables, describe the variable relationship between latent variables and observed variables, and construct a structural equation model.

[0118] The expression of the variable relationship in step S421 is: in, Indicates the RMS value of vibration, Indicates mechanical wear, Indicates the SF6 gas concentration, Indicates insulation degradation, Indicates the amount of partial discharge, Indicates environmental corrosion, represents the rate of temperature rise, Indicates the Factor loading coefficients, Indicates the measurement error term.

[0119] The above factor loading coefficient is used to quantify the explanatory power of latent variables on observed variables, such as =0.85 means that mechanical wear explains 85% of the vibration RMS variation, and the measurement error term follows a normal distribution, representing sensor noise or unmodeled factors.

[0120] The expression of the structural equation model in step S421 is:

[0121] in, Indicates the path coefficients, Indicates the A structural residual term.

[0122] In step S421, the present invention first defines latent variables and observed variables, describes the variable relationship between these latent variables and observed variables, and constructs a structural equation model. In power distribution room equipment status analysis, latent variables are often the root causes of equipment failure. This invention uses mechanical wear, insulation degradation, and environmental corrosion as the root causes. These latent variables are abstract concepts that cannot be directly measured, while observed variables are specific indicators measured by various sensors. This invention uses vibration RMS value, SF6 gas concentration, partial discharge, and temperature rise rate as specific indicators. These are directly measurable physical quantities.

[0123] S422: Calculate the path coefficients between variables of the structural equation model through partial least squares path analysis.

[0124] The expression of the path coefficient between variables in step S422 is:

[0125]

[0126] in, represents the estimated path coefficient between variables, Indicates the total number of paths, Indicates the Insulation degradation caused by the Indicates the Mechanical wear caused by the path, Indicates the Environmental corrosion caused by a path.

[0127] Partial least squares path analysis is used to calculate the path coefficients between variables in the structural equation model. Partial least squares path analysis (PLS-PA) is an estimation method that does not rely on the normal distribution assumption and is suitable for situations where the sample size is small and there is multicollinearity between variables. PLS-PA estimates model parameters by maximizing the covariance between latent variables. The iterative process includes external model estimation, internal model estimation, and weight update, and finally outputs the path coefficients between variables, such as , , it indicates that mechanical wear is the main driving factor for insulation degradation.

[0128] S423: Based on the structural equation model and the path coefficients between the variables, root cause verification is performed through a counterfactual intervention method to obtain root cause analysis results.

[0129] The counterfactual intervention method is a root cause analysis method based on causal reasoning. It evaluates the causal effect of each factor on the result by simulating counterfactual scenarios of how the result would change if a certain factor was changed.

[0130] Within the framework of structural equation models, counterfactual interventions can be implemented by adjusting specific model parameters or exogenous variables and then recalculating the model's predicted values. For example, to verify whether mechanical wear is the root cause of abnormal vibration in a piece of equipment, one can set mechanical wear to 0 in the model, indicating no wear, and then predict the RMS vibration value. If the predicted value is significantly lower than the actual observed value, this supports the view that mechanical wear is the root cause of the abnormal vibration. Similarly, by systematically conducting intervention experiments on various potential root causes and comparing the magnitude of the intervention effects, the root cause with the greatest impact on the abnormality can be identified.

[0131] S424: Output the equipment abnormality determination result and the root cause analysis result to obtain a health status assessment result of the distribution room.

[0132] Finally, the equipment anomaly determination and root cause analysis results are output to obtain the distribution room health status assessment results. The health status assessment results include information such as the current equipment status (normal / abnormal), the anomaly type (sequential anomaly / sudden anomaly), and the root cause analysis results. This provides a comprehensive basis for subsequent operation and maintenance decisions. For example, the final output of the distribution room health status assessment results shows that an anomaly is detected in circuit breaker No. 1 (meeting both sequential anomaly and sudden anomaly conditions). The anomaly indicators are the vibration RMS value and temperature rise rate. The root cause analysis shows that mechanical wear is the primary cause. Based on this primary cause, it is recommended to inspect the circuit breaker transmission mechanism and consider replacing key components based on their service life.

[0133] S5: Making an operation and maintenance decision based on the health status assessment result of the power distribution room, and performing operation and maintenance on the power distribution room according to the operation and maintenance decision.

[0134] In step S5, an operation and maintenance decision is made based on the health status assessment result of the distribution room, and the distribution room is operated and maintained through the operation and maintenance decision. This means that based on the health status assessment result of the distribution room obtained in the previous steps, a specific operation and maintenance strategy is generated through the operation and maintenance decision algorithm, and then the corresponding operation and maintenance operations are executed. This can achieve a closed loop from state perception to action execution, and convert the results of anomaly detection and root cause analysis into operational operation and maintenance actions.

[0135] The generation of O&M decisions begins with a risk assessment of abnormal equipment identified in the health status assessment results. Based on the type of anomaly (sequential or sudden), severity, and root cause analysis, anomalies are categorized into three risk levels: urgent, major, and general. Urgent risks are anomalies that could cause immediate equipment failure or endanger the safety of the distribution room, such as mechanically stuck circuit breakers or transformer winding shorts. Major risks are anomalies that could develop into serious failures in the short term, such as poor contact in switchgear or persistently high transformer oil temperatures. General risks are anomalies that require attention but do not immediately impact equipment operation, such as minor vibration anomalies or slow temperature fluctuations.

[0136] Next, for anomalies of varying risk levels, corresponding O&M strategies are generated based on a pre-defined decision rule base. This decision rule base includes a mapping between device type, anomaly pattern, root cause type, and recommended O&M measures. Specific O&M measures include: immediate shutdown for maintenance, live inspection, planned maintenance, enhanced condition monitoring, and routine inspections. For urgent risks, immediate shutdown for maintenance is typically adopted; for major risks, live inspection or scheduling of near-term maintenance can be selected; for general risks, planned maintenance or enhanced monitoring is implemented.

[0137] The generated O&M strategies are then prioritized and resource allocated. Prioritization takes into account equipment importance, risk level, and resource availability. Equipment importance is determined by its location and function in the distribution system; for example, main transformers are typically considered the most important. Resource allocation involves the scheduling and allocation of maintenance personnel, spare parts, and maintenance tools to ensure smooth O&M execution.

[0138] Finally, specific operation and maintenance operations are performed according to the operation and maintenance strategy, including on-site inspection, equipment testing, parts replacement, parameter adjustment, etc. After the operation and maintenance is completed, effect verification is required to check whether the abnormal status has been improved to ensure the effectiveness of the operation and maintenance actions.

[0139] In a specific embodiment, for example, in a circuit breaker mechanical failure case, during preliminary data collection, the vibration sensor on circuit breaker No. 1 in a 10kV distribution room detected a sharp increase in the vibration RMS value from the normal level of 1.0 to 2.8, which lasted for 48 hours. At the same time, the temperature sensor recorded that the temperature rise rate in the dynamic mechanism area increased from the normal value of 1.0 to 1.3, and the current transformer detected that the tripping time was extended from the standard 80ms to 120ms.

[0140] When detecting time series anomalies and sudden anomalies through the isolation forest-long short-term memory hybrid model of the present invention, the anomaly scores were 0.85 and 0.82, respectively, both higher than the threshold; root cause analysis was performed through the structural equation model, and the main root cause was determined to be mechanical wear of the operating mechanism, with an attribution confidence of 0.92. Specifically, the friction force increased due to aging of the transmission bearing. The evaluation result was ultimately a major risk that needed to be handled in a timely manner, otherwise it might develop into a circuit breaker refusal to operate failure.

[0141] Based on the health status assessment results, the operation and maintenance decision generation system outputs the following operation and maintenance strategy: arrange for maintenance personnel to conduct live infrared and acoustic inspections on circuit breaker No. 1 within 72 hours to confirm the fault location; plan a power outage for maintenance next Tuesday (a low-load period) to replace the operating mechanism bearings and perform mechanical lubrication; increase the monitoring frequency before maintenance, and record the change trend of tripping time every four hours. This strategy takes into account the importance of the equipment, the risk of failure, and the system operating conditions, and balances equipment safety and power supply reliability.

[0142] like Figure 2 As shown, the present invention also provides a smart distribution room operation and maintenance system based on multi-source data fusion, including:

[0143] Acquisition module 100: used to collect distribution room environment and equipment data based on multi-source sensing equipment to obtain original multi-source data;

[0144] Standardization module 200: used to perform standardization processing on the original multi-source data to obtain standardized heterogeneous data;

[0145] Feature extraction module 300: used to perform spatiotemporal correlation modeling of device vibration characteristics and device power parameters through a spatiotemporal graph convolutional network, and to fuse the cross-modal deep features extracted from the standardized heterogeneous data to obtain panoramic perception features;

[0146] Detection module 400: configured to perform anomaly detection using an isolation forest-long short-term memory hybrid model based on the panoramic perception features, and deploy a causal inference engine to analyze causal relationships between multiple variables to obtain a health status assessment result of the distribution room;

[0147] Decision output module 500: used to make operation and maintenance decisions based on the health status assessment results of the distribution room, and operate and maintain the distribution room based on the operation and maintenance decisions.

[0148] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0149] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0150] The present invention provides a smart distribution room operation and maintenance method and system based on multi-source data fusion. By modeling the spatiotemporal correlation between equipment vibration characteristics and power parameters using a spatiotemporal graph convolutional network, combined with a hybrid anomaly detection model and a causal reasoning engine, it achieves accurate identification and root cause location of distribution equipment anomalies, with significant technical effects. The present invention can not only simultaneously detect timing anomalies and sudden anomalies, but also accurately trace the root cause of the anomalies, providing a comprehensive and accurate decision-making basis for distribution room operation and maintenance, significantly improving the pertinence and effectiveness of preventive maintenance, reducing unnecessary power outages and maintenance, extending equipment service life, reducing operation and maintenance costs, and improving the overall reliability and safety of the distribution system.

[0151] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A smart distribution room operation and maintenance method based on multi-source data fusion, characterized in that: include: S1: Collect distribution room environment and equipment data through multi-source sensing equipment to obtain original multi-source data; S2: performing standardization processing on the original multi-source data to obtain standardized heterogeneous data; S3: The spatiotemporal correlation between the equipment vibration characteristics and the equipment power parameters is modeled through a spatiotemporal graph convolutional network, and the cross-modal deep features extracted from the standardized heterogeneous data are integrated to obtain panoramic perception features; Wherein, step S3 further includes: S31: Based on the connection relationship of the equipment in the distribution room, establish an equipment topology map, adjust the edge weights of the equipment topology map according to the operating status data of the equipment, and obtain a dynamic adjacency matrix; S32: input the dynamic adjacency matrix into the spatiotemporal graph convolutional network to obtain a spatiotemporal joint feature vector; S33: perform attention interaction calculation on the equipment power parameters and the spatiotemporal joint feature vector to obtain a coupling coefficient matrix, and perform abnormal area detection based on the threshold segmentation method on the coupling coefficient matrix to obtain coupling features; S34: perform feature extraction on the standardized heterogeneous data to obtain cross-modal deep features; S35: fuse the coupling features and the cross-modal deep features to obtain panoramic perception features; S4: Based on the panoramic perception features, anomaly detection is performed using an isolation forest-long short-term memory hybrid model. A causal inference engine is deployed to analyze the causal relationship between multiple variables to obtain a health status assessment result for the distribution room. Wherein, step S4 further includes: S41: Inputting the panoramic perception features into an isolation forest-long short-term memory hybrid model to obtain equipment anomaly determination results including time series anomalies and sudden anomalies; S42: Using a structural equation model, root causes of abnormal devices in the equipment anomaly determination results are located to obtain a distribution room health status assessment result including equipment anomaly determination results and root cause analysis results; Wherein, step S42 further includes: S421: defining latent variables and observed variables, describing the variable relationships between latent variables and observed variables, and constructing a structural equation model; S422: calculating the path coefficients between variables of the structural equation model through partial least squares path analysis; S423: performing root cause verification through counterfactual intervention method based on the structural equation model and the path coefficients between variables to obtain root cause analysis results; S424: outputting the equipment abnormality determination result and the root cause analysis result to obtain the health status assessment result of the distribution room; S5: Making an operation and maintenance decision based on the health status assessment result of the power distribution room, and performing operation and maintenance on the power distribution room according to the operation and maintenance decision.

2. The intelligent distribution room operation and maintenance method based on multi-source data fusion according to claim 1 is characterized in that: The original multi-source data in step S1 includes: Ambient temperature and humidity data, which is collected by a temperature and humidity sensor; Gas concentration data, the gas concentration data is collected by a gas concentration sensor; Power operation parameter data, the power operation parameter data is collected by a power parameter collector; Equipment vibration characteristic data, the equipment vibration characteristic data is collected by a vibration sensor; Thermal imaging data, which is collected by infrared thermal imaging equipment.

3. The intelligent distribution room operation and maintenance method based on multi-source data fusion according to claim 1 is characterized in that: Step S2 specifically includes: Performing noise removal and outlier detection on the original multi-source data using a data cleaning algorithm to obtain cleaned data; Normalizing the numerical data in the cleaned data to the maximum and minimum values through normalization processing to obtain normalized numerical data; Performing a logarithmic transformation on the skewed distribution data in the cleaned data to obtain balanced distribution data; The time synchronization mechanism is used to align the timestamps of data with different sampling frequencies to obtain time-series consistent data; Completing the missing parts in the time series consistent data by using a missing value filling algorithm to obtain integrity-enhanced data; The integrity enhanced data is converted into a unified format through a data format converter to obtain the standardized heterogeneous data.

4. The intelligent distribution room operation and maintenance method based on multi-source data fusion according to claim 1 is characterized in that: Step S32 further includes: S321: performing spectral graph convolution on the dynamic adjacency matrix through Chebyshev polynomial approximation to extract spatial correlation features of the equipment vibration spectrum; S322: Inputting the spatial correlation features into a gated temporal convolutional network to extract temporal correlation features of the vibration signal time series; S323: combining the spatial correlation features and the temporal correlation features to obtain fusion features; S324: Add the fused features to the original vibration signal in the standardized heterogeneous data through residual connection to obtain a spatiotemporal joint feature vector.

5. The intelligent distribution room operation and maintenance method based on multi-source data fusion according to claim 1 is characterized in that: Step S41 specifically includes: S411: Calculating prediction residuals using the LSTM branch of the isolation forest-long short-term memory hybrid model, marking devices whose prediction residuals are greater than a first preset threshold as abnormal, and obtaining a first marking result; The expression of the first preset threshold is: t >3σ, where ∈ t is the forecast residual, σ is the standard deviation of the residual series; S412: Calculating a path length score using the isolation forest branch of the isolation forest-long short-term memory hybrid model, and marking devices with a path length score greater than a second preset threshold as abnormal to obtain a second marking result; The expression of the second preset threshold is: S f >0.8, where S f Score path length; S413: When either the first marking result or the second marking result is marked as abnormal, the device abnormality determination result is output as abnormal; when either the first marking result or the second marking result is marked as no abnormality, the device abnormality determination result is output as normal.

6. The intelligent distribution room operation and maintenance method based on multi-source data fusion according to claim 1 is characterized in that: The expression of the variable relationship in step S421 is: Among them, V rms Indicates the RMS value of vibration, L mech Indicates mechanical wear, C SF6 Indicates SF6 gas concentration, L ins Indicates insulation degradation, Q pd Indicates the amount of partial discharge, L env Indicates environmental corrosion, represents the temperature rise rate, λ k represents the kth factor loading coefficient, where k = 1, 2, ..., 6, δ e represents the e-th measurement error term, where e = 1, 2, …, 4; The expression of the structural equation model in step S421 is: Among them, γ i represents the i-th path coefficient, where i=1,2,3, ξ o represents the oth structural residual term, where o = 1, 2; The expression of the path coefficient between the variables in step S422 is: in, represents the estimated path coefficient between variables, n represents the total number of paths, represents the insulation degradation caused by the i-th path, represents the mechanical wear caused by the i-th path, represents the environmental corrosion caused by the i-th path.

7. A smart distribution room operation and maintenance system based on multi-source data fusion, characterized in that: include: Acquisition module: used to collect distribution room environment and equipment data based on multi-source sensing equipment to obtain original multi-source data; Standardization module: used for standardizing the original multi-source data to obtain standardized heterogeneous data; Feature extraction module: used to perform spatiotemporal correlation modeling of equipment vibration characteristics and equipment power parameters through a spatiotemporal graph convolutional network, and to fuse the cross-modal deep features extracted from the standardized heterogeneous data to obtain panoramic perception features; The feature extraction module is specifically used to: establish a device topology map based on the device connection relationship in the distribution room, adjust the edge weights of the device topology map according to the device operating status data, and obtain a dynamic adjacency matrix; input the dynamic adjacency matrix into the spatiotemporal graph convolutional network to obtain a spatiotemporal joint feature vector; perform attention interaction calculation on the device power parameters and the spatiotemporal joint feature vector to obtain a coupling coefficient matrix, and perform abnormal area detection on the coupling coefficient matrix based on the threshold segmentation method to obtain coupling features; perform feature extraction on the standardized heterogeneous data to obtain cross-modal deep features; Fusing the coupling feature and the cross-modal depth feature to obtain a panoramic perception feature; Detection module: used to detect anomalies using an isolation forest-long short-term memory hybrid model based on the panoramic perception features, and deploy a causal inference engine to analyze the causal relationship between multiple variables to obtain a health status assessment result of the distribution room; The detection module is specifically configured to: input the panoramic perception features into an isolation forest-long short-term memory hybrid model to obtain an equipment anomaly determination result including a time series anomaly and a sudden anomaly; locate the root cause of the abnormal equipment in the equipment anomaly determination result using a structural equation model to obtain a distribution room health status assessment result including an equipment anomaly determination result and a root cause analysis result; The detection module is further configured to: define latent variables and observed variables, describe the variable relationships between latent variables and observed variables, and construct a structural equation model; calculate the path coefficients between variables of the structural equation model through partial least squares path analysis; perform root cause verification through a counterfactual intervention method based on the structural equation model and the path coefficients between variables to obtain root cause analysis results; output the equipment abnormality determination result and the root cause analysis result to obtain a health status assessment result of the distribution room; Decision output module: used to make operation and maintenance decisions based on the health status assessment results of the distribution room, and operate and maintain the distribution room based on the operation and maintenance decisions.

Images